Non-dogmatic social discounting
By Antony Millner∗
The long-run social discount rate has an enormous effect on
the value of climate mitigation, infrastructure projects, and
other long-term public policies. Its value is however highly
contested, in part because of normative disagreements about
social time preferences. I develop a theory of ‘non-dogmatic’
social planners, who are insecure in their current norma-
tive judgments and entertain the possibility that they may
change. Although each non-dogmatic planner advocates an
idiosyncratic theory of intertemporal social welfare, all such
planners agree on the long-run social discount rate. Non-
dogmatism thus goes some way towards resolving normative
disagreements, especially for long-term public projects.
JEL: H43,D61,D90
Keywords: Social discount rate, normative uncertainty, inter-
dependence, cost-benefit analysis.
‘I take the problem of discounting for projects with payoffs in the far fu-
ture...to be largely ethical.’ – Kenneth Arrow (1999)
∗ Department of Economics, University of California, Santa Barbara, CA 93106, USA,[email protected]. I am grateful to Geir Asheim, Partha Dasgupta, Marc Fleurbaey,Simone Galperti, Ben Groom, Geoff Heal, Derek Lemoine, Lucija Muehlenbachs, FrikkNesje, Bruno Strulovici, the audiences of numerous seminars and conferences, four anony-mous referees, and the editor, for helpful comments and discussions. This work was car-ried out at the Grantham Research Institute, London School of Economics and PoliticalScience. I gratefully acknowledge financial support from the ESRC Centre for ClimateChange Economics and Policy and the Grantham Foundation for the Protection of theEnvironment during my time at LSE.
1
2 THE AMERICAN ECONOMIC REVIEW
‘[T]he list of axioms we use as a basis for our ethical theory can never be
more than a tentative list, one always open to possible revision.’
– John Harsanyi (1977)
The social discount rate (SDR) converts the future consequences of public
projects into present values, and is thus a critical input to public cost-
benefit analysis. Small changes in the SDR can have an enormous effect on
the estimated value of public projects with long-run consequences such as
infrastructure investments, climate change mitigation measures, and nuclear
waste management (Arrow et al., 2013). Yet despite almost a century of
economic research on intertemporal public decision-making, opinion is still
divided on how costs and benefits that occur more than a few decades in
the future should be discounted.
SDRs are related to intertemporal marginal rates of substitution, which
quantify the social value of consumption changes in the future. Much of
the disagreement about SDRs stems from normative disagreements about
the social time preferences that determine these marginal rates of substitu-
tion (Drupp et al., 2018a; Dasgupta, 2008). That such disagreements occur
should not be surprising – specifying social time preferences requires dif-
ficult normative judgments about, for example, the appropriate degree of
social impatience and aversion to intertemporal consumption inequalities,
and there is no silver bullet specification that is immune to criticism (see
Greaves (2017) for a discussion of the arguments). Conversely, as MacAskill
(2016) observes, ‘for almost any ethical view, there seems to be something
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 3
to be said in its favour.’1
This paper develops a theory of ‘non-dogmatic’ social planners’ time pref-
erences, starting from the premise that no single normative theory of in-
tertemporal social welfare is unassailable, and devotees of all theories should
thus exhibit a degree of insecurity in their normative judgments.2 Although
non-dogmatic planners favour a particular theory of intertemporal social
welfare, they admit the possibility that they may be persuaded of the virtues
of an alternative theory in the future. Non-dogmatic planners anticipate
these possible changes, and internalize the preferences of their future selves.
The non-dogmatism of such planners thus manifests both through their will-
ingness to entertain the possibility of a change in their views, and through
their unwillingness to act as pure dictators with respect to their future
selves.3 Since non-dogmatic planners are always insecure, their normative
preferences at any future time τ reflect uncertainty about future preferences
at times greater than τ . Persistent normative insecurity coupled with inter-
nalization of future preferences thus results in a recursively defined sequence
of time preferences, in which current planners’ preferences depend on their
1An alternative tradition in the literature, often termed the ‘positive approach’, iden-tifies SDRs with observed market interest rates, and thus does not directly engage withnormative reasoning. Arrow, Dasgupta and Maler (2003) however remind us that ‘us-ing market observables to infer social welfare can be misleading in imperfect economies.That we may have to be explicit about welfare parameters...in order to estimate marginalrates of substitution in imperfect economies is not an argument for pretending that theeconomies in question are not imperfect after all.’ Market imperfections are particularlysalient for long-run SDRs. Gollier and Hammitt (2014), for example, explain that ‘thepositive approach cannot be applied for time horizons exceeding 20 or 30 years, becausethere are no safe assets traded on markets with such large maturities.’
2Throughout the paper I use ‘planner’ and ‘theory’ roughly interchangeably. A plan-ner’s time preferences are equivalent to a normative theory of intertemporal social welfare.They are thus distinct from consumer preferences that are inferred by revealed preference,but are an intertemporal analogue of the social welfare functions used in, for example,optimal tax theory.
3Non-dogmatic planners still rank consumption streams using their current prefer-ences, but current preferences depend in part on future preferences.
4 THE AMERICAN ECONOMIC REVIEW
possible future preferences, each of which is in turn recursively defined.
Crucially, non-dogmatic planners still make idiosyncratic judgments about
all the contested normative aspects of intertemporal welfare functions (IWFs),4
including utility functions and pure time discount factors. Nevertheless, I
show that all non-dogmatic IWFs yield the same value of the long-run SDR.
Thus, adopting a model of intertemporal evaluation in which planners ex-
hibit some insecurity in their normative judgments ends up resolving dis-
agreements about the evaluation of long-run public projects. The intuition
for this finding is developed in an example below. It is a consequence of the
fact that each non-dogmatic IWF depends in part on other non-dogmatic
IWFs in future periods. Disagreements about long-run SDRs wash out when
this nested sequence of interdependent valuations is unravelled backwards
to the present, since IWFs mix repeatedly over time.
Formally, the model extends and reinterprets existing models of ‘purely
altruistic’ intergenerational time preferences (Ray, 1987; Bergstrom, 1999;
Saez-Marti and Weibull, 2005; Galperti and Strulovici, 2017). In these mod-
els a single representative agent in the current generation internalizes the
preferences of future generations, assuming that each future generation does
the same, and that preferences are time invariant. Leading philosophers have
long drawn an analogy between present generations’ concerns for future
generations, and present selves’ concerns for future selves.5 Parfit (1984,
p. 319), for example, argues that ‘Like future generations, future selves
have no vote, so their interests need to be specially protected.’ The present
4IWFs are functions that represent planners’ normative preferences over consumptionstreams.
5Analogies between intergenerational and intrapersonal choice have also proved fruit-ful in economics (Phelps and Pollak, 1968; Laibson, 1997). See Ray, Vellodi and Wang(2018) for a behavioural model in which agents exhibit concern for future selves.
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 5
paper formalizes this analogy in a model of social planners whose norma-
tive judgments may change over time. Unlike the intergenerational work
cited above, the internalization of future preferences occurs intrapersonally
in my model.6 Hori (2009) has previously observed in a static framework
that increased interdependence between heterogeneous agents who internal-
ize others’ preferences can cause their values to get ‘closer together’. This
paper shows that in any fixed model of heterogeneous social planners’ time
preferences internalization leads to complete convergence on the most con-
tested quantity in public cost-benefit analysis: the long-run SDR.
A related literature has studied social choice theoretic approaches to ag-
gregating time preferences7(Gollier and Zeckhauser, 2005; Heal and Millner,
2014; Millner and Heal, 2018; Chambers and Echenique, 2018; Feng and
Ke, 2018), and the aggregation of heterogeneous opinions on SDRs (Weitz-
man, 2001; Freeman and Groom, 2015).8 Unlike this work, this paper does
not specify an aggregation rule that is applied unilaterally by an external
analyst. Non-dogmatic planners may disagree on all the normative issues
that are sources of contention in discussions of social discounting, and priv-
ilege their own theory of intertemporal social welfare. Nevertheless, non-
dogmatism causes each planner to account to some extent for alternative
theories, so that some aggregation occurs internally in each theory. I show
that this is enough to generate consensus on the long-run SDR.
6An alternative interpretation of the model that retains an interpersonal flavour ispossible, see footnote 15.
7A utilitarian aggregation approach leads to representative discount rates that aredominated by the preferences of the most patient agent for long maturities (Gollier andZeckhauser, 2005). This only occurs in a very special case of the model I develop; ingeneral consensus long-run discount rates are determined by a non-trivial mixture of allnon-dogmatic planners’ IWFs.
8The papers that aggregate SDRs directly do not disentangle heterogeneous beliefsabout facts (i.e., consumption growth rates) from disagreements about values. Thispaper, like the social choice literature, focusses on values.
6 THE AMERICAN ECONOMIC REVIEW
It is important to emphasize that the model I present is normative: I sug-
gest that planners should exhibit some insecurity in their normative judg-
ments, propose a method for them to do so, and use a calibrated version
of the model to show how observed disagreements on SDRs might change
if they did. The model does not claim to describe the observed behaviour
of governments, or the recommendations of ‘experts’ on social discounting.
Like much normative work, the paper is an exercise in persuasion. It shows
that if advocates of alternative theories of intertemporal social welfare ad-
mitted some insecurity in their normative judgments, but were unwilling to
give them up entirely, a lot of progress could still be made.
I. A motivating example
The essential features of the model can be illustrated in a simple ex-
ample. Suppose that there are only two plausible normative theories of
intertemporal social welfare, and let planner i ∈ {1, 2} be a devotee of the-
ory i. To establish a baseline model, begin by assuming that at time τ
planner i’s normative preferences over infinite annual consumption streams
Cτ = (cτ , cτ+1, cτ+2, . . .) can be represented by an IWF V iτ of the following
familiar form:
(1) V iτ = U i(cτ ) + βiV
iτ+1,
where U i(c) is a utility function and βi ∈ (0, 1) is a pure time discount
factor. These IWFs have the following equivalent representation:
(2) V iτ =
∞∑s=0
(βi)sU i(cτ+s).
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 7
Now consider evaluating a marginal public project with a sequence of annual
payoffs πππ = (π0, π1, . . .). Standard results (Dasgupta, Sen and Marglin,
1972; Gollier, 2012) show that the project is welfare improving according
to planner i if and only if its net present value is positive, where the net
present value of πππ is defined as:
(3) NPV i(πππ) =∞∑s=0
πse−ri(s)s,
and the social discount factor at maturity s is given by the marginal rate
of substitution between consumption at times τ + s and τ , denoted MRSis:
(4) e−ri(s)s = MRSis = (βi)
s (U i)′(cτ+s)
(U i)′(cτ ).
The social discount rate at maturity s according to planner i is
(5) ri(s) = −1
slnMRSis.
This fundamental quantity tells us how planner i converts safe marginal
payoffs at maturity s to present values. Since each planner has an idiosyn-
cratic utility function U i(c) and pure time discount factor βi, there is no
possibility of them generically agreeing on any part of the term structure of
SDRs ri(s).
Equation (5) can be made more intuitive by assuming that utility func-
tions are iso-elastic (i.e., (U i)′(c) = c−ηi), writing βi = e−ρi , and defining
compound annual consumption growth rates gs via cτ+s = cτegss. Substi-
8 THE AMERICAN ECONOMIC REVIEW
tuting these assumptions into (4–5) we find the famous Ramsey rule:
(6) ri(s) = ρi + ηigs.
The first term in this expression is planner i’s pure rate of social time pref-
erence, and the second term captures her aversion to intertemporal con-
sumption inequalities, which depends on her elasticity of marginal utility ηi.
Planners’ adopted values for these parameters constitute primitive norma-
tive judgments about how society should trade off consumption at different
points in time (Gollier and Hammitt, 2014).
Now consider a variation on the time preferences in (1). Suppose that
each planner is a little insecure in her normative judgments, and entertains
the possibility that she may be persuaded of the alternative theory of in-
tertemporal social welfare in the future. For concreteness, suppose that the
probability of the planners’ judgments remaining unaltered next year is w,
and the probability of them changing is 1 − w. How should the planners
account for their insecurity today? One answer is that they should simply
forget about it. This is perfectly coherent, but amounts to a dogmatic impo-
sition of current preferences on future selves, despite the planners’ insecurity
about their current, possibly transitory, normative judgments. Normative
insecurities have no consequences for SDRs in this case.9 A second option is
for the planners to adjust their ‘raw’ preferences by aggregating them with
the alternative theory of intertemporal social welfare. However, this places
9This is the approach often taken in models of time inconsistent preferences. Sophis-ticated agents in these models anticipate the actions of their future selves, and reactoptimally to them, but do not incorporate future preferences into their own rankings ofconsumption streams – they are dogmatic. See Galperti and Strulovici (2017) for furtherdiscussion of the relationship between time consistency and preference internalization.
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 9
current and possible future preferences on an equal conceptual footing today,
even though the planners are currently devotees of only one theory. We are
after a model in which current planners can put all their eggs in one basket.
A third option – the one I will pursue – is for the planners to use their cur-
rent preferences to rank consumption streams, but for those preferences to
internalize the preferences of future selves. In this interpretation equation
(1) is thought of as saying that self τ ’s IWF is an additive combination of
current utility and the IWF of self τ + 1. If the self at τ admits the possi-
bility that the self at τ + 1 may be persuaded of the alternative theory of
intertemporal social welfare, a natural analogue of (1) is:
V 1τ = U1(cτ ) + β1(wV 1
τ+1 + (1− w)V 2τ+1),(7)
V 2τ = U2(cτ ) + β2((1− w)V 1
τ+1 + wV 2τ+1),(8)
where w ∈ (0, 1).
In this model planners’ insecurity in their current normative judgments
causes them to avoid imposing their current preferences on their future self
(they only care about the self one year ahead in this example). Current
planners account for their future self’s preferences directly, and do not just
dogmatically value future consumption streams using their current prefer-
ences, which may be obsolete by the time next year rolls around. Moreover,
normative insecurity is persistent: planners’ preferences at time τ + 1 them-
selves reflect uncertainty about preferences at τ + 2, ad infinitum. Planners
whose IWFs are of the form in (7–8) will be called ‘non-dogmatic’ – I provide
a formal definition below.10 Note that the IWFs defined by (7–8) still admit
10The literature on intergenerational altruism uses the terms ‘non-paternalistic’ or‘pure’ to describe agents who internalize others’ preferences. I use ‘non-dogmatic’ both
10 THE AMERICAN ECONOMIC REVIEW
arbitrary idiosyncratic pure time discount factors and utility functions.
To analyze the coupled system of time preferences in (7–8), define
~Vτ =
V 1τ
V 2τ
; ~Uτ =
U1(cτ )
U2(cτ )
; F =
β1w β1(1− w)
β2(1− w) β2w
.
Then we can write (7–8) as:
(9) ~Vτ = ~Uτ + F~Vτ+1 =∞∑s=0
Fs~Uτ+s.
Planners’ attitudes to consumption changes in the distant future depend on
the behaviour of Fs for large s. Since w ∈ (0, 1), the matrix F is strictly
positive. The Perron-Frobenious theorem (see Sternberg, 2014) then tells
us that there is a 2× 2 matrix A, with elements aij > 0, such that
(10) lims→∞
Fs
µs= A,
where µ ∈ (0, 1) is the largest eigenvalue of F. Thus when s is large both
planners’ weights on future utilities are proportional to µs, where µ is a
non-trivial mixture of both planners’ discount factors.11
To understand the intuition for this result notice that current planners at
τ only care about utilities at future times τ+1, τ+2, . . . indirectly through a
mixture F of their possible IWFs at τ+1. Planners at τ+1 in turn only care
about utilities at times τ +2, τ +3, . . . through a mixture F of their possible
to distinguish my model of intrapersonal internalization from this literature, and becausethis term is a better fit to the context of this paper, in which planners contend with manyplausible theories of intertemporal welfare.
11In this example µ = w(β1+β2)2 +
√w2(β1+β2)2
4 − β1β2(2w − 1).
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 11
IWFs at time τ +2. Thus we see that current planners’ attitudes to utilities
at time τ+s are obtained by iterating their possible IWFs at τ+s backwards
to τ , passing through their IWFs at times τ+s−1, τ+s−2, . . . , τ+1. With
each step back in this iteration the discount factors associated with different
theories of intertemporal welfare are mixed by the matrix F. As the number
of mixing operations grows (i.e., as s increases), planners’ discount factors
become homogenized. For large s the mixing process converges, and both
planners’ long-run utility weights are proportional to a common factor µs.
It is the fact that planners anticipate possible changes in their theories of
intertemporal social welfare, and form their current preferences with one eye
on their future selves, that delivers this result.
Substituting (10) into (9) we see that according to planner i, the marginal
rate of substitution between consumption at τ and consumption at distant
future times τ + s is
MRSis =µs[ai1(U1)′(cτ+s) + ai2(U2)′(cτ+s)]
(U i)′(cτ ).(11)
With a few assumptions we can simplify this expression further. Denote the
long-run growth rate of consumption by g, i.e., cτ+s = egscτ for large s. In
addition, define the long-run pure rate of social time preference ρ = − lnµ,
and assume again that utility functions are iso-elastic (i.e., (U i)′(c) = c−ηi).
Since (U i)′(cτ+s) ∝ e−gηis for large s, MRSis is dominated by the exponential
term with the lowest value of ηig. Substituting these assumptions into (11),
we see that when s is large,
MRSis ∝ e−(ρ+min{η1g,η2g})s ⇒ ri(s)→ ρ+ min{η1g, η2g}.
12 THE AMERICAN ECONOMIC REVIEW
Thus, although the non-dogmatic planners may have arbitrary disagree-
ments about pure time discount factors and elasticities of marginal utility,
they both agree on the long-run SDR. I will show below that disagreements
may reduce substantially even for medium term maturities.
II. The model
I now extend the results above to an arbitrary number of planners, each
of whom may account for the preferences of future selves into the indefinite
future. Assume that there are N > 1 plausible normative theories of in-
tertemporal social welfare. As before I identify planner i with theory i, and
denote i’s IWF at time τ by V iτ . The vector of IWFs at time τ is denoted
by ~Vτ = (V 1τ , V
2τ , . . . , V
Nτ ). We will say that the time preferences defined by
the sequence {~Vτ}τ∈N are non-dogmatic if for all i = 1 . . . N, τ ∈ N,
V iτ = U i(cτ ) +
∞∑s=1
N∑j=1
f ijs Vjτ+s,(12)
where f ijs ≥ 0 for all s ≥ 1, and there exists a t ≥ 1 such that f ijt > 0 for all
i, j = 1 . . . N . Lemma 1 in the Appendix shows that (12) defines a unique
bounded set of time preferences that are non-decreasing in all utilities if
(13) maxi
{∞∑s=1
N∑j=1
f ijs
}< 1.
I assume this condition from now on.12
The definition in (12) encodes three assumptions. First, planners’ time
12The condition in (13) is sufficient, but not necessary, for the required properties tohold. A necessary and sufficient condition is provided in the appendix, however as thiscondition is difficult to check in practice we will work with (13). None of the resultsdepend on this simplification.
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 13
preferences are forward looking and time invariant; this captures the persis-
tence of normative insecurity, and implies that preferences do not depend on
the history of consumption. Second, current planners internalize the prefer-
ences of possible future selves, and assign non-zero weight to each plausible
theory when imagining what their future preferences might be. Third, pref-
erences are additively time separable. A set of IWFs satisfies these three
assumptions if and only if it is of the form in (12).13
The intertemporal weight f ijs in (12) is the product of two terms: the pure
time discount factor of planner i at time τ on the IWF of self τ +s (denoted
by βis), and the weight i places on theory j in year τ + s (denoted by wijs ):
f ijs = βiswijs ,
where∑N
j=1wijs = 1. In the normative application we consider it is natural
to require some parity between the weights wijs of different planners, as in
the simple example above.14 This ensures that the model delivers a set of
theories that are ‘equally non-dogmatic’, but since this is not required for
the main result I do not insist on it in the definition.
There is nothing in the representation (12) that requires us to think of the
weights wijs as probabilities – at present these weights are merely parameters
of the preference representation.15 If, however, we do interpret these weights
13Forward looking time invariant IWFs that internalize future preferences are of theform V iτ = F i(cτ , V
1τ+1, . . . , V
Nτ+1, V
1τ+2, . . . , V
Nτ+2, . . .). Galperti and Strulovici (2017)
show that IWFs of this kind are time separable if and only if they are of the form in (12).14Equation A.32 in the appendix gives a more sophisticated example of ‘parity’ between
planners’ intertemporal weights.15With some modification (12) could be interpreted as a positive model of a set of
altruistic agents, each of whom cares about everyone else’s total wellbeing in every futureperiod. This would require the arguments of utility functions to be idiosyncratic privateconsumption variables, rather than aggregate social consumption – in this case each agentwould have N consumption discount rates at each maturity. If, however, these agents
14 THE AMERICAN ECONOMIC REVIEW
as beliefs, it is natural to require that that those beliefs be consistent. Con-
sistency requires that current planners’ beliefs about which theories they
may adopt in the future cohere with their future selves’ beliefs about their
own chances of switching from their preferred theory. Let Probτ (i → j; s)
denote the probability that planner i at time τ assigns to a switch to theory
j after s years. Beliefs are consistent iff
Probτ (i→ j; s) =N∑k=1
Probτ (i→ k; t)Probτ+t(k → j; s− t),
for all τ ∈ N, s ≥ 2, 1 ≤ t < s. Lemma 2 in the Appendix shows that non-
dogmatic planners have consistent beliefs iff there exists an N×N stochastic
matrix P such that
(14) wijs = (Ps)i,j,
for all s ≥ 1. We use this restriction on the weights wijs in Section IV below,
but the main results do not require it.
III. Results
As in the example above, V iτ in (12) has an equivalent representation in
terms of sums of future utilities which may be determined by solving the
infinite system of equations (12) (see appendix). We write the solution of
this system as
(15) V iτ =
∞∑s=0
N∑j=1
aijs Uj(cτ+s),
derived their utility from consumption of a public good, (12) would apply unchanged.
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 15
where aijs ≥ 0 for all i, j = 1 . . . N, s ∈ N. The SDR at maturity s according
to planner i is:
(16) ri(s) = −1
slnMRSis = −1
sln
(1
(U i)′(cτ )
N∑j=1
aijs (U j)′(cτ+s)
).
Define the elasticity of planner i’s marginal utility function as
(17) ηi(c) = −c(U i)′′(c)
(U i)′(c).
If ηi(c) is uniformly larger than ηj(c), planner i is more averse to intertem-
poral consumption inequalities than planner j. I assume that ηi(c) ≥ 0, is
bounded for all c, and that limc→∞ ηi(c) > 0 and limc→0 η
i(c) > 0 for all i
(I assume that all limits exist). In addition, define the long-run growth rate
of consumption to be
g = lims→∞
1
sln
(cτ+s
cτ
)and let
(18) η =
mini {limc→∞ ηi(c)} if g > 0
maxi {limc→0 ηi(c)} if g < 0.
Finally, let Fs be an N × N matrix with elements (Fs)i,j = f ijs , let 1N be
16 THE AMERICAN ECONOMIC REVIEW
the N ×N identity matrix, and define the NM ×NM matrix
ΦM =
F1 F2 . . . FM−1 FM
1N 0 . . . 0 0
0 1N . . . 0 0...
......
......
0 0 . . . 1N 0
.(19)
Let µ(M) ∈ (0, 1) be the largest eigenvalue of ΦM .
With these definitions in place the main result can be stated.
PROPOSITION 1: All non-dogmatic planners agree on the long-run SDR:
(20) lims→∞
ri(s) = ρ+ ηg, ∀i = 1, . . . , N,
where ρ = − limM→∞ lnµ(M).
The proof of this proposition shows that the requirement in (12) that each
planner place positive weight on all theories in some future period is stronger
than is needed for the result.16 The formula in (20) can also be extended
to the case where consumption growth is uncertain (see the appendix). The
proposition also provides a practical procedure for approximating ρ: com-
pute − lnµ(M) for increasingly large values of M .17
Proposition 1 provides a simple characterization of the consensus long-
run elasticity of marginal utility η. The consensus long-run pure rate of
social time preference ρ is, however, a much more complex quantity, which
16It is sufficient for each planner to place positive weight on some other theory in somefuture period, in such a way that if we look far enough ahead, all planners’ preferencesinfluence each other. Planner i need not place positive weight on theory j directly.
17The appendix shows that − lnµ(M) decreases monotonically to ρ as M increases.
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 17
depends on the full set of intertemporal weights f ijs . The appendix provides
further discussion of ρ, including some comparative statics results. We will
content ourselves with describing two intuitive properties of ρ here.
PROPOSITION 2: 1) ρ is decreasing in f ijs for all i, j, s.
2) Suppose that the intertemporal weights f ijs are given by
f ijs (ε) =
f iis j = i
hijs (ε) j 6= i,
where the functions hijs (ε) are continuous, hijs (ε) > 0 for ε > 0, and
hijs (0) = 0. Let ρi be planner i’s idiosyncratic long-run rate of pure
time preference when ε = 0, and let ρ(ε) be the consensus long-run
rate of pure time preference when ε > 0. Then
(21) limε→0+
ρ(ε) = miniρi.
The first part of the proposition is intuitive. Any increase in f ijs increases the
weight planner i places on future utilities. Since all planners’ IWFs depend
on planner i’s IWF, all planners are less impatient if f ijs increases. Thus the
consensus long-run rate of time preference decreases if f ijs increases. The
second part of the proposition shows that if all planners assign arbitrarily
small, but positive, weight to alternative theories, their consensus long-
run rate of time preference is the lowest of all of their ‘dogmatic’ rates.
To understand the intuition for this finding, note that although planner i
places arbitrarily small weight on theories that do not coincide with her
current theory as ε → 0, each theory still enters into her current IWF V iτ
for all ε > 0. When ε = 0 planner j’s weights on future utilities decline
18 THE AMERICAN ECONOMIC REVIEW
like e−ρjs as s → ∞. Thus the planner with the lowest value of ρj will
place exponentially more weight on distant future utilities than any more
impatient planner as s → ∞ when ε = 0. Since the most patient planner’s
preferences are part of each planner’s preferences for ε > 0, by continuity
the consensus long-run rate of time preference must be given by the most
patient planner’s dogmatic long-run rate of time preference as ε→ 0.18
Part 2 of Proposition 2 invokes related findings on the aggregation of
opinions on SDRs (Weitzman, 2001; Freeman and Groom, 2015), and on
the utilitarian aggregation of time preferences (e.g. Gollier and Zeckhauser,
2005). In each of these cases averaging over a distribution of discount factors
leads to a ‘certainty equivalent’ discount rate, or a representative discount
rate, that declines to the lowest rate as the time horizon tends to infinity.
Proposition 2 differs from these results as it pertains to the long-run SDR
in each theory, rather than an external analysts’ average across preferences
or real discount rates. The proposition also shows that ρ is only determined
by the most patient planner in a very special case of the model, i.e., when
planners are ‘minimally’ non-dogmatic. In all other cases, ρ is a non-trivial
mixture of the intertemporal weights of all theories.
IV. Consequences for cost-benefit analysis
While Proposition 1 emphasizes the emergence of a consensus on the
long-run SDR, this result implies a more general phenomenon that has
18More technically, a matrix that determines planners’ pure time discount factors,call it Φ(ε), separates into N independent components at ε = 0, each of which has adominant eigenvalue that corresponds to the long-run pure time discount factor of oneof the dogmatic planners. The largest eigenvalue of Φ(0) is simply the largest of theseN dominant eigenvalues. When ε > 0, Φ(ε) is primitive and has only one component.Since eigenvalues are continuous functions of matrix elements, the largest eigenvalue ofΦ(ε) converges to the largest of the N dogmatic long-run discount factors as ε→ 0.
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 19
relevance for cost-benefit analysis. As (3) shows, calculations of the net
present value of public projects depend on the full term structure of SDRs.
Since non-dogmatic planners’ SDRs ri(s) converge completely as maturity
s → ∞, they must also exhibit partial convergence at finite maturities.
Non-dogmatism may thus reduce disagreement about project NPVs by act-
ing through the entire term structure of the SDR. In this section, I illustrate
the effect of non-dogmatism on cost-benefit analysis of public projects in a
calibrated numerical model, and also show how quickly disagreements about
SDRs may decline with maturity. The results in this section constitute a
normative counterfactual: I have argued that advocates of all theories of
intertemporal welfare should be non-dogmatic, and this section illustrates
what might happen to disagreements if they were.
To enable this analysis I will work with data on economists’ opinions on
the normatively appropriate IWF, collected by Drupp et al. (2018a,b).19 Al-
though there is no deep reason why economists’ normative judgments should
be seen as representative of the distribution of plausible theories, they do
arguably have an advantage in understanding the quantitative implications
of different recommendations for cost-benefit analysis. Rawls (1971), in his
notion of ‘reflective equilibrium’, argues that this is an essential feature
of good normative reasoning. For my purposes these economists’ opinions
merely provide an interesting distribution of informed views on these mat-
ters.
19I assume that the opinions expressed in the survey data do not already account fornon-dogmatism. Drupp et al. (2018a) state explicitly that ‘we structure the survey arounda well-known framework for inter-temporal welfare evaluations: Time Discounted Utili-tarianism’, and work with an iso-elastic utility function. Only two respondents objectedto the survey’s request for a constant pure rate of time preference, and none objected tothe request for a constant elasticity of marginal utility (personal communication). Bothof these quantities would vary with maturity if the respondents were non-dogmatic.
20 THE AMERICAN ECONOMIC REVIEW
The Drupp et al. (2018a) survey contains 173 complete responses from
scholars who have published papers on social discounting. Each respondent
gave an opinion on the appropriate values of the parameters of a discounted
utilitarian IWF with iso-elastic utility function. The 5-95% ranges of opin-
ions on the pure rate of social time preference and elasticity of marginal
utility were [0,3.85%/yr] and [0.2,3] respectively. To calibrate the model I
assume that the intertemporal weights f ijs = βiswijs in (12) take the following
form:
βis = γi(αi)s, wijs = (Ps)i,j, where Pi,j =
x i = j
1−xN−1
i 6= j.(22)
The parameter x ∈ [0, 1] is the probability that planners stick to their cur-
rent theory next year.20 Conditional on switching, planners assign an equal
probability to all other theories. By (14), planners’ beliefs are consistent
in this model. The values of γi and αi are calibrated so that when x → 1
non-dogmatic planners’ IWFs are a close approximation to a discounted
utilitarian IWF, and consistent with the values for the pure rate of social
time preference that survey respondents reported. Utility functions U i(c)
are taken to be iso-elastic, with the elasticity of marginal utility calibrated
to respondents’ reported values.
One subtlety of the calibration procedure is worth pointing out. I have
chosen to present the model with an annual time step as discount rate
20In this model, wiis = x(Nx−1N−1
)s−1
+ 1N
(1−
(Nx−1N−1
)s−1)
for all i, and wijs =1−wii
s
N−1
for j 6= i. However, the appendix explains that the numerical results in this section arerobust to alternative specifications of the weights wijs for s ≥ 2. Note that these areweights on future IWFs, and not on future utilities; utility weights are determined by thesolution of the entire system (12). No matter what IWF planner i adopts in year s, it isdiscounted using her current intertemporal weight γi(αi)
s.
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 21
schedules are usually presented at this temporal resolution in practice. If,
however, we change the time step we also have to change the values of the
dynamical parameters in the model, i.e., consumption growth rates, rates of
pure time preference, and in particular the transition probability matrix P,
to reflect the change of units. This procedure is less straightforward in this
model of interdependent preferences than in more familiar dynamic models,
but can be accomplished. The appendix provides further details on this
point, and a full description of the calibration procedures.
Given this calibration methodology, the distribution of term structures for
the SDR can be computed for different values of the parameter x. Figure 1a
depicts the results of this exercise, assuming a constant consumption growth
rate of 2%/yr. The figure shows that disagreements about the SDR could
reduce substantially even at medium term maturities if planners were non-
dogmatic. Reductions in disagreement are greatest at longer maturities, but
are substantial even for maturities of 30 years. When planners’ judgments
are highly persistent (i.e., x is close to 1) the range of opinions on SDRs
expands, but for any x < 97.5% disagreement is reduced by more than
a factor of three at maturities greater than 50 years. In the appendix I
demonstrate that the rapid reduction in disagreement depicted in Figure 1a
is largely driven by non-dogmatic planners’ elasticities of marginal utility
(i.e., the analogue of the consumption growth term ηgs in (6)), and not by
their rates of pure social time preference (i.e., the analogue of ρ in (6), which
depends only on the intertemporal weights f ijs ). Disagreements about the
consumption growth term in the Ramsey rule are significantly larger than
disagreements about the pure rate of time preference, but decay rapidly with
maturity if planners are non-dogmatic. Disagreements about the pure rate
22 THE AMERICAN ECONOMIC REVIEW
of time preference are smaller, but decay much more slowly with maturity.
More than 90% of the variation in r(s) is attributable to variation in the
pure rate of time preference alone for maturities s > 60 years, for all the
values of x depicted in Figure 1a.
Figure 1b illustrates how non-dogmatism reduces disagreements about the
net present values of payoff sequences, as defined by (3). The figure depicts
five payoff sequences πππ.21 To quantify the reduction in disagreement about
NPVs let σ({NPV (πππ;x)}) denote the standard deviation of the set of net
present values of πππ according to non-dogmatic planners with weight x in
(22), and compute the following ratio for each sequence πππ:
(23) Γ(πππ;x) =σ({NPV (πππ;x)})σ({NPV (πππ; 1)})
.
This ratio captures the reduction in disagreement about NPVs, relative to
the dogmatic benchmark at x = 1. Figure 1b shows that non-dogmatism
could substantially reduce disagreements about the value of projects whose
payoffs are concentrated at maturities of 30 years or greater, even if x =
97.5%. Reductions in disagreement increase strongly as payoffs move further
into the future. For the project on the far right, whose benefits largely occur
more than 60 years in future, disagreements are reduced by more than a
factor of 5 even if x = 97.5%.
The rate of dissipation of disagreement with maturity depicted in Fig-
ure 1 clearly depends on the value of the parameter x. The values x =
80%, 90%, 95%, 97.5% used in this figure correspond to a change of norma-
21These may be thought of as public projects with equal up-front costs but differ-ent temporal profiles of benefits. The undiscounted sum of benefits for each project isnormalized to 1.
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 23
0 20 40 60 80 100 120 140 160 180 200
Maturity s (yrs)
0
1
2
3
4
5
6
7
5-9
5%
range o
f r
i (s)
(%/y
r)
x = 80%
x = 90%
x = 95%
x = 97.5%
x = 1
(a) Simulated 5-95% range for non-dogmatic planners’ SDRs ri(s). The solid black curve
corresponds to x = 1 in (22), dotted purple x = 97.5%, dash-dotted yellow x = 95%, dashed
red x = 90%, and solid blue x = 80%. Consumption growth is a constant 2%/yr.
0 20 40 60 80 100 120Years (s)
Proj
ect p
ayof
fs (
s)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(b) Reduction in disagreement about NPVs when planners are non-dogmatic. Each curve in
the figure denotes a time sequence of payoffs. The markers centered on each curve denote thevalues of Γ(πππ, x), defined in (23), for this payoff sequence. ◦,+,×, � denote values of Γ(πππ, x)when x = 97.5%, 95%, 90%, 80% respectively in (22).
Figure 1. : Consequences of non-dogmatism for cost-benefit analysis.
24 THE AMERICAN ECONOMIC REVIEW
tive views roughly once every 5, 10, 20, and 40 years respectively, on average.
Are these plausible values? In order to answer this question we must first
recognise that x can be interpreted either as a positive or a normative ob-
ject. On the one hand, any ethical observer’s insecurity in their normative
judgments could be construed as a subjective matter; in this interpretation
assessing the chance of those views changing is a positive question about
that observer’s state of mind. There is some suggestive evidence for the oc-
casional changes of heart that are needed to support the paper’s conclusions
under this interpretation, as professional philosophers’ convictions have been
shown to correlate with their age (Bourget and Chalmers, 2014).22 On the
other hand if we accept the persuasive motivation for the model, i.e., that it
is designed to nudge planners into forming their normative judgments in a
new way, then x plays a normative role. Tweaking this lever shows planners
how their normative views on IWFs should adjust away from the standard
discounted utilitarian framework, given a degree of non-dogmatism that is
‘normatively required’. Although both interpretations are consistent with
the model, the latter is more in keeping with the ethos of this paper.
Given this, it is reasonable to ask how much non-dogmatism is ‘norma-
tively required’. That is something of a meta-ethical question, and readers
will doubtless have their own views on it. Requiring planners to admit the
possibility of a change in their convictions roughly once every 10 or 20 years
does not seem like an excessively burdensome prescription (recall that they
still have the freedom to discount the preferences of future selves as they see
fit). Indeed, the argument that uncertainty or insecurity should play a non-
22This finding chimes with a witticism often attributed to Georges Clemencau: ‘Notto be a socialist at twenty is proof of want of heart; to be one at thirty is proof of wantof head.’
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 25
negligible role in normative judgments is not unique to this paper. Catholic
theology has grappled with related issues since the 16th century, when the
doctrine of ‘probabilism’ was introduced as a guide to action in the face of
moral uncertainty (Harty, 1913). Normative uncertainty is currently also
a central topic in philosophy, precisely because many have grown weary of
old debates that pit ethical theories against one another in a zero-sum game
(see e.g. Bostrom, 2009; MacAskill, 2016; MacAskill and Ord, 2018).
V. Conclusion
This paper introduced a normative model of social planners’ time pref-
erences based on a principle of ‘non-dogmatism’. This principle requires
advocates of alternative theories of intertemporal social welfare to exhibit
a degree of humility when forming their normative judgments: They admit
the possibility of a change in their views, and refrain from imposing their
current normative judgments on their future selves. The formalism allows
advocates of each theory the freedom to express idiosyncratic judgments
about all the contested normative aspects of social time preferences. In
spite of this, all non-dogmatic theories yield the same value of the long-run
social discount rate. As the appropriate value of this quantity has been
widely contested and has a powerful influence on the evaluation of pub-
lic projects with long-run consequences, this analysis may prove useful for
policy applications.
REFERENCES
Arrow, Kenneth J. 1999. “Discounting, morality, and gaming.” In Discounting and
Intergenerational Equity. , ed. Paul R. Portney and John P. Weyant. Resources for the
Future.
26 THE AMERICAN ECONOMIC REVIEW
Arrow, Kenneth J., Partha Dasgupta, and Karl-Goran Maler. 2003. “Evaluating
Projects and Assessing Sustainable Development in Imperfect Economies.” Environ-
mental and Resource Economics, 26(4): 647–685.
Arrow, K., M. Cropper, C. Gollier, B. Groom, G. Heal, R. Newell, W. Nord-
haus, R. Pindyck, W. Pizer, P. Portney, T. Sterner, R. S. J. Tol, and M.
Weitzman. 2013. “Determining Benefits and Costs for Future Generations.” Science,
341(6144): 349–350.
Bergstrom, Theodore C. 1999. “Systems of Benevolent Utility Functions.” Journal of
Public Economic Theory, 1(1): 71–100.
Bostrom, Nick. 2009. “Moral uncertainty - towards a solution?” http: // www.
overcomingbias. com/ 2009/ 01/ moral-uncertainty-towards-a-solution. html ,
accessed on 20/05/2019.
Bourget, David, and David J. Chalmers. 2014. “What do philosophers believe?”
Philosophical Studies, 170(3): 465–500.
Chambers, Christopher P., and Federico Echenique. 2018. “On Multiple Discount
Rates.” Econometrica, 86(4): 1325–1346.
Dasgupta, Partha. 2008. “Discounting climate change.” Journal of Risk and Uncer-
tainty, 37(2): 141–169.
Dasgupta, Partha, Amartya Sen, and Stephen A Marglin. 1972. Guidelines for
project evaluation. New York:United Nations.
Drupp, Moritz, Mark C. Freeman, Ben Groom, and Frikk Nesje. 2018a. “Dis-
counting Disentangled.” American Economic Journal: Economic Policy, 10(4): 109–
134.
Drupp, Moritz, Mark C. Freeman, Ben Groom, and Frikk Nesje. 2018b. “Dis-
counting Disentangled: Dataset.” American Economic Journal: Economic Policy,
https://doi.org/10.1257/pol.20160240.
Feng, Tangren, and Shaowei Ke. 2018. “Social Discounting and Intergenerational
Pareto.” Econometrica, 86(5): 1537–1567.
VOL. NO. NON-DOGMATIC SOCIAL DISCOUNTING 27
Freeman, Mark C., and Ben Groom. 2015. “Positively Gamma Discounting: Com-
bining the Opinions of Experts on the Social Discount Rate.” The Economic Journal,
125(585): 1015–1024.
Galperti, Simone, and Bruno Strulovici. 2017. “A Theory of Intergenerational Al-
truism.” Econometrica, 85(4): 1175–1218.
Gollier, Christian. 2012. Pricing the Planet’s Future: The Economics of Discounting
in an Uncertain World. Princeton University Press.
Gollier, Christian, and James K. Hammitt. 2014. “The Long-Run Discount Rate
Controversy.” Annual Review of Resource Economics, 6(1): 273–295.
Gollier, Christian, and Richard Zeckhauser. 2005. “Aggregation of Heterogeneous
Time Preferences.” Journal of Political Economy, 113(4): 878–896.
Greaves, Hilary. 2017. “Discounting for public policy: A survey.” Economics and Phi-
losophy, 33(3): 391–439.
Harsanyi, John C. 1977. “Rule utilitarianism and decision theory.” Erkenntnis,
11(1): 25–53.
Harty, John M. 1913. “Probabilism.” In Catholic Encyclopedia, Volume 12. The Ency-
clopedia Press.
Heal, Geoffrey M., and Antony Millner. 2014. “Agreeing to disagree on climate
policy.” Proceedings of the National Academy of Sciences, 111(10): 3695–3698.
Hori, Hajime. 2009. “Nonpaternalistic altruism and functional interdependence of social
preferences.” Social Choice and Welfare, 32(1): 59–77.
Laibson, David. 1997. “Golden Eggs and Hyperbolic Discounting.” The Quarterly Jour-
nal of Economics, 112(2): 443–478.
MacAskill, William. 2016. “Normative Uncertainty as a Voting Problem.” Mind,
125(500): 967–1004.
MacAskill, William, and Toby Ord. 2018. “Why Maximize Expected Choice-
Worthiness?” Nous, , (doi: 10.1111/nous.12264): 1–27.
28 THE AMERICAN ECONOMIC REVIEW
Millner, Antony, and Geoffrey Heal. 2018. “Discounting by committee.” Journal of
Public Economics, 167: 91–104.
Parfit, Derek. 1984. Reasons and Persons. Oxford University Press, USA.
Phelps, E. S., and R. A. Pollak. 1968. “On Second-Best National Saving and Game-
Equilibrium Growth.” The Review of Economic Studies, 35(2): 185–199.
Rawls, John. 1971. A theory of justice. Harvard University Press.
Ray, Debraj. 1987. “Nonpaternalistic intergenerational altruism.” Journal of Economic
Theory, 41(1): 112–132.
Ray, Debraj, Nikhil Vellodi, and Ruqu Wang. 2018. “Backward discounting.”
Working paper.
Saez-Marti, Maria, and Jorgen W. Weibull. 2005. “Discounting and altruism to
future decision-makers.” Journal of Economic Theory, 122(2): 254–266.
Sternberg, Shlomo. 2014. Dynamical Systems. Dover Publications.
Weitzman, Martin L. 2001. “Gamma Discounting.” The American Economic Review,
91(1): 260–271.
Non-dogmatic Social Discounting
Online Appendix
Antony Millner∗1
1Department of Economics, University of California, Santa Barbara.
Contents
A Proof of Lemma 1 2
B Proof of Lemma 2 3
C Proof of Proposition 1 4
D Consensus long-run SDRs under uncertainty 14
E Proof of Proposition 2 16
F Comparative statics of the consensus long-run pure rate of social time
preference 17
G Details of calibration 20
H Changing the model’s time step 23
I Decomposing non-dogmatic SDRs 26
J References 31
∗Millner: [email protected].
1
A Proof of Lemma 1
Lemma 1. The system (12) defines a unique bounded set of time preferences, which are
non-decreasing in all utilities, if
maxi
{∞∑s=1
N∑j=1
f ijs < 1
}.
Proof. The system of time preferences (12) can be written as a single matrix equation as
follows:
V 1τ...
V Nτ
V 1τ+1...
V Nτ+1...
=
U1(cτ )...
UN(cτ )
U1(cτ+1)...
UN(cτ+1)...
+
~0N f 111 . . . f 1N
1 f 112 . . . f 1N
2 . . ....
......
......
......
...
~0N fN11 . . . fNN1 fN1
2 . . . fNN2 . . .
~0N ~0N f 111 . . . f 1N
1 f 112 . . . . . .
......
......
......
......
~0N ~0N fN11 . . . fNN1 fN1
2 . . . . . ....
......
......
......
...
V 1τ...
V Nτ
V 1τ+1...
V Nτ+1...
where ~0N is an 1×N vector of zeros. Letting ~Xτ denote the vector on the left hand side
of this expression, Λ the infinite dimensional square matrix on the right hand side, and ~Uτ
denote the vector of Us on the right hand side, we have
~Xτ = ~Uτ + Λ ~Xτ
⇒ ~Xτ = (1∞ −Λ)−1~Uτ ,
where 1∞ is the infinite dimensional identity matrix, and we have assumed that the relevant
matrix inverse exists.
In general infinite dimensional matrices do not have unique inverses. However, Lemma
1 in Bergstrom (1999) shows that 1∞−Λ has a unique bounded inverse with non-negative
elements if and only if 1∞ − Λ is a dominant diagonal matrix. A denumerably infinite
matrix 1∞ − Λ with Λ ≥ 0 is said to be dominant diagonal if there exists a bounded
diagonal matrix D ≥ 0 such that the infimum of the row sums of (1∞ −Λ)D is positive.
Clearly, a sufficient condition for 1∞−Λ to be dominant diagonal is if∑∞
s=1
∑Nj=1 f
ijs < 1
for all i.
2
Although this lemma focusses on providing a sufficient condition that is easy to check,
the proof also provides a necessary and sufficient condition: 1∞ − Λ must be dominant
diagonal. This is equivalent to requiring the spectral radius of the linear operator Λ to be
less than 1, as this guarantees that the sequence (1∞−Λ)−1 = 1∞+Λ+Λ2 + . . . converges
(Duchin & Steenge, 2009). Checking this condition is however difficult in practice given
the infinite dimensionality of Λ. I will thus work with the simpler sufficient condition
throughout, but the results do not depend on this simplification. The proof of the main
proposition in Appendix C only requires the spectral radius of Λ to be bounded above by
1.
B Proof of Lemma 2
We wish to prove that non-dogmatic planners’ with preferences (12) have consistent beliefs
iff the intratemporal weights wijs satisfy (14). In the notation established in the text,
Lemma 2.
Probτ (i→ j; s) =N∑k=1
Probτ (i→ k; t)Probτ+t(k → j; s− t) (A.1)
for all τ ∈ N, s ≥ 2, 1 ≤ t < s if and only if there exists an N × N stochastic matrix P
such that
wijs = (Ps)i,j.
Let the beliefs of planners at time τ about the probability of a future self who subscribes
to theory i at time τ+s−1 switching to theory j at time τ+s be Tij,(τ)s . Denote the matrix
of these transition probabilities by T(τ)s . Let W
(τ)s be the matrix of time τ planners’ beliefs
about which theory they will subscribe to at time τ+s, whose i, j element is Probτ (i→ j; s).
Then we have
W(τ)s = T(τ)
s T(τ)s−1 . . .T
(τ)1 .
Using this relation, (A.1) can be written as the requirement that
T(τ)s T
(τ)s−1 . . .T
(τ)1 = T
(τ+t)s−t T
(τ+t)s−t−1 . . .T
(τ+t)1 T
(τ)t T
(τ)t−1 . . .T
(τ)1 , (A.2)
for all τ, t, s. It is clear that a sufficient condition for this to be satisfied is
T(τ)s = P
3
for all τ, s, where P is an N ×N stochastic matrix. To prove necessity, put s = 2, t = 1 in
(A.2) to find
T(τ)2 T
(τ)1 = T
(τ+1)1 T
(τ)1
which implies
T(τ)2 = T
(τ+1)1 . (A.3)
Putting s = 3, t = 1 in (A.2), we find
T(τ)3 T
(τ)2 T
(τ)1 = T
(τ+1)2 T
(τ+1)1 T
(τ)1
⇒T(τ)3 T
(τ)2 = T
(τ+1)2 T
(τ+1)1 .
and using (A.3) this reduces to
T(τ)3 = T
(τ+1)2 .
Repeating this process of substitution, we find that a necessary condition for (A.2) to be
satisfied is
T(τ)s+1 = T(τ+1)
s .
Since non-dogmatic planners’ preferences are time invariant, it must be the case that
T(τ+1)s = T(τ)
s .
Substituting this relation into the previous equation shows that
T(τ)s+1 = T(τ)
s
for all τ, s. This implies that the matrix of planners’ beliefs W(τ)s must be of the form
W(τ)s = (P)s
for all τ .
C Proof of Proposition 1
We prove a more general version of the result in Proposition 1. The proof has two main
steps. First we find conditions under which all planners’ utility weights aijs are proportional
to a common discount factor µs for large s. We then show that when these conditions are
4
satisfied all non-dogmatic planners’ long-run SDRs are the same.
STEP 1:
Begin by defining the sequence of N ×N matrices
Fs :=
f 11s f 12
s . . . f 1Ns
f 21s f 22
s . . . f 2Ns
......
......
fN1s fN2
s . . . fNNs
(A.4)
and the sequences of N × 1 vectors
~Vτ =
V 1τ
V 2τ...
V Nτ
, ~Uτ =
U1(cτ )
U2(cτ )...
UN(cτ )
. (A.5)
Our general model (12) can be written as:
~Vτ = ~Uτ +∞∑s=1
Fs~Vτ+s. (A.6)
We seek an equivalent representation of this system of the form
~Vτ :=∞∑s=0
As~Uτ+s, (A.7)
where As is a sequence of N ×N matrices of the form,
As :=
a11s a12
s . . . a1Ns
a21s a22
s . . . a2Ns
......
......
aN1s aN2
s . . . aNNs
(A.8)
where aijs is the weight planner i at time τ assigns to theory j’s utility function at time
τ + s, i.e., U j(cτ+s).
We now prove the following:
Proposition A.I. Assume that the condition (13) is satisfied, and that f iis > 0 for all
5
i = 1 . . . N , s = 1 . . .∞. Construct a directed graph G with N nodes labelled 1, 2, . . . , N .
Draw an edge from node i to node j 6= i iff f ijs > 0 for at least one s ≥ 1. If G contains a
directed cycle of length N , then there exists a µ ∈ (0, 1) such that
lims→∞
aijsµs
= Kij > 0
where the Kij are finite constants.
Notice that the definition of non-dogmatic time preferences in (12) automatically im-
plies that the directed cycle condition in this proposition is satisfied (the graph G is com-
plete in this case, i.e., all edges exist). However, the directed cycle condition itself is
considerably weaker than is assumed in this definition.
Proof. Substitute (A.7) into (A.6) to find
∞∑s=0
As~Uτ+s = ~Uτ +
∞∑p=1
Fp
(∞∑q=0
Aq~Uτ+p+q
)(A.9)
Equating coefficients of ~Uτ+s in this expression, we see that As must satisfy
A0 = 1N (A.10)
As =s∑
p=1
FpAs−p for s > 0. (A.11)
where 1N is the N×N identity matrix. The solution of this recurrence relation determines
the utility weights aijs . It will be convenient to split this matrix recurrence relation into a
set of N vector recurrence relations as follows. Let ~Ajs be the j-th column vector of As,
i.e.,
~Ajs =
a1js
a2js...
aNjs
. (A.12)
6
Define ~ej to be the unit vector with elements
(~ej)i =
{0 i 6= j
1 i = j
Then (A.11) is equivalent to the N vector recurrence relations
~Aj0 = ~ej
~Ajs =s∑
p=1
Fp~Ajs−p for s > 0. (A.13)
for j = 1 . . . N .
The proof now has the following steps. We consider finite order models, i.e., FM ′ = 0 for
all M ′ greater than some finite M . We show that if a certain augmented matrix constructed
from the matrices F1, . . . ,FM is primitive, all planners will have a common long-run pure
time discount factor. A square matrix B is primitive if there exists an integer k > 0 such
that Bk > 0. We then extend this result to infinite order models by taking an appropriate
limit of finite order models. Finally, we show that primitivity of the required matrices in
the infinite order case is ensured by the graph theoretic condition in the statement of the
proposition.
Begin with the finite order case. Let M = max{s|∃i, j f ijs > 0} <∞. In this case, for
all s > M , (A.13) reduces to
~Ajs =M∑p=1
Fp~Ajs−p. (A.14)
Define the NM ×NM matrix
ΦM =
F1 F2 . . . FM−1 FM
1N 0 . . . 0 0
0 1N . . . 0 0...
......
......
0 0 . . . 1N 0
(A.15)
7
where 1N is the N ×N identity matrix. In addition, define the ‘stacked’ vector
~Y js =
~Ajs~Ajs−1
...~Ajs−M+1
Then we can rewrite the Mth order recurrence (A.14) as a first order recurrence as follows:
~Y js = ΦM
~Y js−1
⇒ ~Y jM+s = (ΦM)s~Y j
M . (A.16)
We now assume that ΦM is a primitive matrix. By the Perron-Frobenius theorem for
primitive matrices (Sternberg, 2014), this implies
1. ΦM has a positive eigenvalue, which we label as µ(M).
2. All other eigenvalues of ΦM have complex modulus strictly less than µ(M).
3. There exists a matrix C > 0 such that
lims→∞
ΦsM
[µ(M)]s= C
4. µ(M) increases when any element of ΦM increases.
5.
µ(M) < maxi
∑j
φij. (A.17)
where φij is the ijth element of ΦM .
Since the first N elements of ~Y js coincide with aijs , the third of these conclusions implies
that
∀i, j, lims→∞
aijs[µ(M)]s
= C~Y jM > 0.
To bound the value of µ(M), note that from point 5 of the Perron-Frobenius theorem
in (A.17), and the definition of ΦM in (A.15), we have
µ(M) < maxi
{M∑s=1
N∑j=1
f ijs
}(A.18)
8
Thus, if∞∑s=1
N∑j=1
f ijs < 1 (A.19)
for all i, µ(M) < 1, and hence lims→∞ aijs = 0. Thus (13) guarantees that the time
preferences (12) are complete (i.e., finite on bounded consumption streams, and hence able
to rank arbitrary pairs of bounded consumption streams) for all finite M . This concludes
the finite M case.
We now extend this result to the case of infinite M . Assume that there exists an M ′ > 0
such that the matrix ΦM , defined in (A.15), is primitive for all M > M ′. For M > M ′,
define
~Vτ (M) = ~Uτ +M∑s=1
Fs~Vτ+s(M)
and let~Vτ = lim
M→∞~Vτ (M).
Define the equivalent representations of these preferences by
~Vτ (M) =∞∑s=0
As(M)~Uτ+s (A.20)
~Vτ =∞∑s=0
As~Uτ+s (A.21)
In addition, let µ(M) be the Perron-Frobenius eigenvalue of ΦM . We begin by proving
that:
Lemma 3.
µ := limM→∞
µ(M) exists. (A.22)
Proof. Consider the eigenvalue µ(M + 1), where M > M ′. This is the Perron-Frobenius
eigenvalue of ΦM+1. The M -th order preferences ~Vτ (M) are equivalent to an M + 1th
order model, with FM+1 = 0. The matrix ΦM , which controls the asymptotic behavior
of ~Vτ (M) can thus be thought of as an N × (M + 1) matrix, where the last M rows
and columns are zeros. Call this matrix ΦM+1. The matrix ΦM+1, associated with the
asymptotic behavior of ~Vτ (M + 1), has entries that are strictly larger than than those of
ΦM+1 in at least some elements. Thus, by point 4 in our statement of the Perron-Frobenius
theorem, µ(M + 1) > µ(M). We also know that µ(M) < 1 for all M . Since the sequence
9
µ(M) is increasing and bounded above, the monotone convergence theorem implies that µ
exists.
We have thus proved that if the matrices ΦM are primitive for M > M ′,
limM→∞
lims→∞
aijs+1(M)
aijs (M)= lim
M→∞µ(M) = µ. (A.23)
Note that since (A.17) and (A.19) are strict inequalities, µ < 1. We now wish to know
whether it is also true that:
lims→∞
limM→∞
aijs+1(M)
aijs (M)= µ. (A.24)
That is, can we change the order of the limits in (A.23)? For limit operations to be
interchangeable we require the sequence of functions they operate on to be uniformly
convergent. The functions in question here are V iτ (M) and V i
τ , which we can think of as
linear functions from the infinite dimensional space R∞ × RN = {(~Uτ , ~Uτ+1, ~Uτ+2, . . .)} to
R. If the sequence of functions V iτ (M) converges uniformly to V i
τ on any bounded subset
of R∞ × RN , then (A.24) will be satisfied. We now prove a second lemma:
Lemma 4. Let B be a compact subset of R∞ × RN , and assume that (13) is satisfied.
Then V iτ (M) converges uniformly to V i
τ on B.
Proof. Equation (A.13) shows that for all s ≤M , aijτ+s(M) = aijτ+s. Let U = maxj{sups{U j(cτ+s)}}be the largest component of any ~U ∈ B. For any ~U ∈ B,
sup~U∈B
∣∣∣V iτ (M)− V i
τ
∣∣∣ = sup~U∈B
∣∣∣∣∣∞∑s=1
N∑j=1
aijτ+M+s(M)U j(cτ+M+s)−∞∑s=1
N∑j=1
aijτ+M+sUj(cτ+M+s)
∣∣∣∣∣≤
∞∑s=1
N∑j=1
[∣∣aijτ+M+s(M)∣∣+∣∣aijτ+M+s
∣∣] UBy Lemma 3, µ < 1 also implies µ(M) < 1 for allM , so we know that limM→∞ a
ijτ+M+s(M) =
0 = limM→∞ aijτ+M+s for all i, j. Thus
limM→∞
sup~U∈B
∣∣∣V iτ (M)− V i
τ
∣∣∣ = 0.
Hence V iτ (M) converges uniformly to V i
τ .
This concludes the infinite order case.
10
The final step of the proof is to show that if the graph G, defined in the statement of
the proposition, has a directed cycle of length N , then there exists an M ′ > 0 such that for
all M > M ′ the matrix ΦM is primitive. We demonstrate this using a graphical argument.
Consider an aribtrary R × R matrix Bij, and form a directed graph H(B) on nodes
1 . . . R, where there is an edge from node i to node j iff Bij > 0. The matrix Bij is primitive
if there exists an integer k ≥ 1 such that there is a path of length k from each node i to
every other node j in H(B). If H(B) is strongly connected, i.e., there exists a path from
every node to every other node, then a sufficient condition for Bij to be primitive is for
there to be at least one node that is connected to itself.
Now consider our NM ×NM matrices ΦM . To construct the directed graph H(ΦM)
associated with ΦM in a convenient form, follow the following procedure: Construct an
M ×N grid of nodes (where N is the number of planners), with node (m,n) representing
planner n at time τ + m. For all m > 1, n, construct a directed edge from node (m,n) to
node (m− 1, n). In addition, construct a directed edge from node (1, n) to node (m′, n′) if
fnn′
m′ > 0.
As an example, take the case M = N = 3, i.e., a third order model with three plan-
ners. In this case ΦM is a 9 × 9 matrix. Assume that f iis > 0 for all i, s = 1 . . . 3, that
f 121 , f 23
1 , f 311 > 0, and that f ijs = 0 otherwise. Figure F.1 represents the directed graph
associated with the matrix Φ3 in this case.
Examination of the figure shows that since f iis > 0, each of the ‘column’ subgraphs
{(m, 1)}, {(m, 2)}, {(m, 3)},m = 1 . . . 3 is strongly connected. Moreover, the cycle between
columns (the red dashed edges) connects the columns to each other, and causes the entire
graph to be strongly connected. Since each node in the first row is connected to itself, the
matrix Φ3 in this example is primitive.
Returning to the general case, suppose that f iis > 0 for all i and s. From the example
in Figure F.1 it is clear that this implies that for each fixed i the subgraph {(m, i)|m =
1 . . .∞} is strongly connected, with each of the nodes (1, i) connected to itself. Thus, if
there is a directed cycle between all of the ‘columns’ of the graph H(ΦM ′) for some M ′,
then for all M > M ′, H(ΦM) is strongly connected, and contains nodes that are connected
to themselves. Hence for all M > M ′, ΦM is a primitive matrix. This concludes the
proof.
STEP 2:
We now show that when the conditions of Proposition A.I are satisfied, all non-dogmatic
theories yield the same long-run SDR, and we compute an explicit formula for this con-
11
Figure F.1: The directed graph H(Φ3) associated with the matrix in our example. Thevertical black edges arise from the identity matrices in the definition of ΦM (see (A.15)).The dashed blue edges arise from f iis > 0, and the dashed red edges from f 12
1 , f 231 , f 31
1 > 0.
sensus discount rate.
Begin by defining
ρ = − ln µ,
where µ is defined in (A.22). When the conditions of Proposition A.I hold we know that
aijs ∼ Kij(s)e−ρs (A.25)
where ∼ denotes asymptotic behaviour as s → ∞, and the multiplicative factors Kij(s)
satisfy lims→∞1s
lnKij(s) = 0.
Now integrate the definition of ηj(c) in (17) to find1
(U j)′(c) = exp
(−∫ c
0
ηj(x)
xdx
).
Make the change of variables x = cτegs′ in the integral in the exponent (recall that g is the
1In other words, solve the differential equation −c(U j)′′/(U j)′ = ηj(c) for (U j)′(c).
12
long-run consumption growth rate), and evaluate (U j)′(c) at c = cτegs to find
(U j)′(cτegs) = exp
(−g∫ s
0
ηj(cτegs′)ds′
).
Defining
ηj =
{limc→∞ η
j(c) g > 0
limc→0 ηj(c) g < 0
(A.26)
we see that the s→∞ asymptotic behaviour of marginal utility is given by
(U j)′(cτegs) ∼ Lj(s)e
−gηjs (A.27)
for some functions Lj(s) that satisfy lims→∞1s
lnLj(s) = 0. Combining (A.25) and (A.27),
we find
ri(s) = −1
sln
(1
(U i)′(cτ )
N∑j=1
aijs (U j)′(cτ+s)
)
∼ −1
sln
(∑j
Kij(s)Lj(s)e−ρse−η
jgs
)
∼ ρ− 1
sln
(∑j
Kij(s)Lj(s)e−ηjgs
)
Define Kij(s) = Kij(s)Lj(s), and let q be the index of the planner with the lowest (highest)
value of ηj when g > 0 (g < 0). Then∑j
Kij(s)Lj(s)e−ηjgs =
∑j
Kij(s)e−ηjgs
= Kiq(s)e−ηqgs
(1 +
∑j 6=q
Kij(s)
Kiq(s)e−(ηj−ηq)gs
)
Since ηj − ηq > 0 for all j 6= q when g > 0, and ηj − ηq < 0 for all j 6= q when g < 0,∑j
Kij(s)Lj(s)e−ηjgs ∼ Kiq(s)e
−ηgs,
13
where η is given by (18). Thus
ri(s) ∼ ρ− 1
sln(Kiq(s)e
−ηgs)
⇒ lims→∞
ri(s) = ρ+ ηg.
D Consensus long-run SDRs under uncertainty
It is straightforward to extend the proof of Proposition 1 to the case where future consump-
tion is uncertain. If consumption is uncertain non-dogmatic planners’ IWFs are simply the
expectation over their deterministic IWFs, i.e.,
V iτ = Ecτ+1,cτ+2,...
∞∑s=0
N∑j=1
aijs Uj(cτ+s)
where Ecτ+1,cτ+2,... denotes the expectation over future consumption values, and the co-
efficients aijs are determined by the dynamical system in (A.11), as in the deterministic
case.
The analysis of the consensus long-run SDR now proceeds in close analogy to the
second part of the proof of Proposition 1. The consensus long-run pure rate of social time
preference is unchanged, however examination of the proof shows that we need to account
for the effect of expectations on the growth terms in the Ramsey rule.
Under uncertainty planners’ marginal rates of substitution between consumption today
and consumption s years from now are given by:
e−ri(s)s = MRSis =
∑Nj=1 a
ijs Ecτ+s(U
j)′(cτ+s)
(U i)′(cτ )(A.28)
Define a planner specific ‘certainty equivalent’ long-run growth rate gj by requiring that
(U j)′(egjscτ ) ≡ Eg(Uj)′(egscτ ) (A.29)
as s→∞, i.e.,
gj ≡ lims→∞
1
slog[((U j)′)−1
(Eg(U
j)′(egscτ ))]. (A.30)
The long-run consumption growth rate g is uncertain in this expression, and Eg denotes
14
expectations over the value of g. In analogy with (A.26), define
ηj(gj) =
{limc→∞ η
j(c) gj > 0
limc→0 ηj(c) gj < 0.
Then for large s, we know from (A.27) that
Ecτ+s(Uj)′(cτ+s) = (U j)′(egjscτ ) ∼ e−gj ηj(gj)s
where ∼ denotes s→∞ asymptotic behaviour, as before.
As in the deterministic case, we see from (A.28) that planner i’s long-run elasticity of
marginal utility is determined by the term that dominates the sum
N∑j=1
aijs Ecτ+s(Uj)′(cτ+s) ∼
∑j
aijs e−gj ηj(gj)s
as s→∞. This sum is dominated by the exponential with the minimum value of gj ηj(gj)
(which may be negative), for all i. We thus conclude that the consensus long-run SDR
under uncertainty is given by
ρ+ mini{giηi(gi)} (A.31)
As an example of the application of this formula suppose that planners’ utility functions
are iso-elastic with elasticities of marginal utility ηi, i.e., (U i)′(c) = c−ηi . In addition,
assume that consumption growth is asymptotically log-normally distributed, i.e.,
log g ∼ N (µ, σ2).
From (A.29) planner i’s certainty equivalent long-run growth rate gi is thus defined by
requiring that at large s,
e−ηigis(cτ )−ηi ≡ Ege
−ηigs(cτ )−ηi = e−(ηiµ− 1
2η2i σ
2)s(cτ )−ηi
⇒ gi = µ− 1
2ηiσ
2
Since elasticities of marginal utility are constant by assumption we know that ηi(gi) = ηi,
and thus the consensus long-run SDR in this example is given by
ρ+ mini{µηi −
1
2η2i σ
2}.
15
E Proof of Proposition 2
Part 1 of the proposition is immediate from point 4 in our statement of the Perron-
Frobenius theorem in Proposition A.I. Part 2 of the proposition follows from the fact
that the eigenvalues of a matrix are continuous in its entries. Consider a set of N ‘dog-
matic’ models, in which each planner assigns weight only to her own theory in future
periods. This set of N independent planners’ time preferences can be represented as a
single non-dogmatic set of N planners as in (12), but where f ijs = 0 if j 6= i. As in the
proof of Proposition A.I, begin by considering a model of finite order M , so that no planner
places any weight on any IWF more than M years ahead. Equation (A.16) shows that the
asymptotic behaviour of such a model can be described by first order difference equations
of the form:~Y js = Φ0
M~Y js−1.
In this case however, the matrix Φ0M , defined in (A.15), is reducible. The largest eigenvalue
of Φ0M is the rate of decline of the utility weights of the most patient dogmatic planner in
the long-run. As M → ∞, the set of eigenvalues of Φ0M contains µi1, the long-run utility
discount factor of planner i, and all eigenvalues of Φ0M are less than or equal to maxi{µi1}.
Now consider the continuous set of models with weights f ijs (ε), where ε > 0. Let
ΦM(ε) be the corresponding ΦM matrix for this set of models, where by assumption
limε→0+ ΦM(ε) = Φ0M . The consensus long-run discount factor in model ε of order M ,
denoted µ1(ε,M) is the largest eigenvalue of ΦM(ε). Define
µ1(ε) = limM→∞
µ1(M, ε).
We know that this limit exists, due to the proof of Proposition A.I. Since the matrix ΦM(ε)
is continuous in ε > 0, and in the limit as M →∞ the largest eigenvalue of ΦM(0) = Φ0M
is equal to maxi{µi1}, we must have
limε→0+
µ1(ε) = maxi{µi1}.
Since ρ(ε) = − ln µ1(ε) by definition, the result follows.
16
F Comparative statics of the consensus long-run pure
rate of social time preference
It is naturally of interest to ask how the consensus long-run pure rate of social time prefer-
ence ρ depends on the intertemporal weights f ijs . Unfortunately strong comparative statics
results on this question are likely out of reach. Technically, we need to understand how the
spectral radius (i.e., largest eigenvalue) of the matrices ΦM from Proposition A.I behaves
when we spread out or contract the distribution of weights f ijs . In order to sign the effect of
a spread in the weights we require something akin to a convexity property for the spectral
radius. Unfortunately, it is known that the spectral radius of a matrix is a convex function
of its diagonal elements, but not of the off-diagonal elements (Friedland, 1981).2
This section describes a special case of the model in which clean comparative statics are
possible. Assume that planner i’s intertemporal weights f ijs depend on a parameter λi ⊂R+, i.e., f ijs = f ijs (λi). Let ~λ = (λ1, . . . , λN) be the vector of planners’ λ parameters, and
assume that ~λ takes values in a convex subset of RN+. Using the notation of Proposition A.I
we write the matrix of weights f ijs at a fixed value of s as Fs(~λ), where we now emphasize
the dependence of these weights on the parameter vector ~λ. We will say that preferences
are symmetric in ~λ iff for all permutation matrices3 P,
Fs(P~λ) = PFs(~λ)PT (A.32)
for all s, where PT is the transpose of P. Intuitively, if preferences are symmetric in~λ, switching any two planners’ values of λ is equivalent to switching their entire set of
intertemporal weights, as this induces a permutation of the weight matrix Fs(~λ). The pa-
rameters λi are thus ‘sufficient statistics’ for planners’ intertemporal weights, and switching
λi ↔ λj is equivalent to relabelling i↔ j.
As an example of preferences that are symmetric in ~λ consider the following:
f ijs =
{β(s, λi)xs j = i
β(s, λi)1−xsN−1
j 6= i(A.33)
2Similarly, it is not possible to sign the effect of premultiplying ΦM by a doubly stochastic matrix, as thespectral radius of a product of two matrices is not sub-multiplicative in general. Gelfand’s formula showsthat the spectral radius of a matrix product is sub-multiplicative if the matrices in question commute, butthis is not much use for our purposes.
3A square matrix is a permutation matrix if each of its rows and each of columns contains exactly oneentry of 1, and zeros elsewhere.
17
where xs ∈ [1/N, 1) for all s = 1 . . .∞, and∑∞
s=1 β(s, λ) < 1 for all λ ∈ I ⊂ R+. In this
model the time dependence of planners’ intertemporal weights f ijs has a common functional
form, given by a discount function β(s, λ) on the IWF of selves s years in the future, where
λ > 0 is a parameter. Variations in planners’ attitudes to time are solely due to differences
in their values of λ. The parametric model defined in (22), which we used in Section IV of
the paper, is of this form if γi = γ for all i.
Let ρ(~λ) be the consensus long-run pure rate of time preference in a model that is
characterized by the parameter vector ~λ.
Proposition A.II. Assume that planners’ time preferences are symmetric in ~λ and that
f ijs (λ) is strictly log-convex in λ > 0 for all i, j, s. Then if the parameter vector ~λA ma-
jorizes4 ~λB,
ρ(~λA) < ρ(~λB).
In words, this result says that if preferences are symmetric in ~λ, intertemporal weights
are log-convex functions of λ, and planners in group A disagree more about the parameter
λ than planners in group B, the consensus long-run pure rate of time preference will be
lower in group A than in group B.
I will provide some interpretation of the log-convexity condition in examples below, but
first we turn to the proof.
Proof. The proof relies on the following result due to Kingman (1961): Let bij(θ) ≥ 0 be
the elements of a non-negative matrix B, where θ ∈ R is a parameter. If bij(θ) is log-
convex in θ for all i, j, the spectral radius of B is a log-convex function of θ. Remark 1.3
in Nussbaum (1986) observes that Kingman’s result can be extended as follows: Let ~θ be
a vector of parameters that takes values in a convex set, and assume that the elements
bij(~θ) ≥ 0 of a matrix B are log-convex functions of ~θ. Then the spectral radius of B is
log-convex is ~θ.
We will employ the usual trick of working with finite order models first (i.e., setting
f ijs to zero for s > M), and taking a limit as M →∞ at the end. The consensus long-run
pure rate of time preference in a model of order M is determined by the largest eigenvalue
of ΦM , defined in (A.15). Denote this eigenvalue by µM(~λ). If the matrix elements f ijs (λ)
are log-convex functions of the scalar variable λ, then f ijs (~λ) = f ijs (λi) is also a log-convex
4~λA majorizes ~λB iff there exists a doubly stochastic matrix H such that ~λB = H~λA. Intuitively, theelements of ~λA are ‘more spread out’ than those of ~λB , and the sums of their elements are equal. Seee.g. Marshall (2010) for a discussion of majorization and its relationship to e.g. stochastic orders andinequality measures.
18
function of the vector of parameters ~λ. Thus, if f ijs (λ) is log-convex (or identically zero)
for all i, j, s, µM(~λ) is a log-convex function of ~λ.
The final step of the proof is to observe that because of the symmetry of the set of
intertemporal weights in (A.32) the spectral radius must be a symmetric function of ~λ, i.e.,
any permutation of the elements of ~λ will leave the spectral radius unchanged. This follows
since the eigenvalues of a matrix are invariant under the permutations (A.32). Since µM(~λ)
is a log convex, symmetric function of ~λ, its log is Schur-convex. Since µM(~λ) = e−ρM (~λ),
this implies that ρM(~λ) is Schur-concave in ~λ. Thus by the properties of Schur-concave
functions, if ~λA majorizes ~λB we must have
ρM(~λA) < ρM(~λB).
The final result follows by taking the limit as M →∞.
As an initial example of the application of this result, consider a model in which the
discount function β(s, λ) in the example in (A.33) declines exponentially, i.e.,
β(s, λ) = (1 + λ)−s .
This discount function satisfies log β(s, λ) = −s log(1 + λ), which is strictly convex in λ.
Thus the result applies – more disagreement about the parameter λ decreases the consensus
long-run pure rate of social time prefenence.
We can extend this finding to a more general class of models by assuming that β(s, λ) =
β(λs), i.e., the parameter λ acts to rescale the time variable s. Following Prelec (2004) we
will say that β(s) exhibits decreasing impatience if log β(s) is a convex function of s for
s > 0. Discount functions that exhibit decreasing impatience have the form β(s) = e−h(s)
where h(s) is a concave function. The rate of increase of h(s) (which measures impatience)
slows as the time horizon s increases.
Corollary 1. Assume that β(s) exhibits decreasing impatience, and that the parameter
vector ~λA majorizes ~λB. Then
ρ(~λA) < ρ(~λB).
Thus, for example, in a hyperbolic model (see e.g. Prelec, 2004) we would have
β(s) = (1 + s)−(1+p) ⇒ β(s, λ) = β(λs) = (1 + λs)−(1+p)
19
where p > 0 is a parameter. β(s) is log convex in s, so more disagreement about λ reduces
the consensus long-run pure rate of time preference in this model.
G Details of calibration
The data I use to calibrate the model and generate the results in Figures 1a and 1b are
taken from a recent survey by Drupp et. al. (2018). They surveyed expert economists who
have published papers on social discounting, asking for their opinions on, amongst other
things, the appropriate values of the pure rate of social time preference and the elasticity
of marginal social utility. The distribution of respondents’ views on these two parameters
is plotted in Figure F.2.
The calibration assumption I use is that the data in Figure F.2 correspond to ‘dogmatic’
views on the IWF, and in particular that these data correspond to the parameters of a
discounted utilitarian IWF with iso-elastic utility function. This assumption is consistent
both with the survey authors’ description of what they aim to elicit in their survey, and
with the participants’ responses. See footnote 19 of the main text for further explanation.
The calibration is made slightly delicate by the fact that there is no version of the
model in (12) in which planners place non-zero weight on all future selves that reduces
to a discounted utilitarian IWF. I calibrate the parametric model in (22) so that when
the weight on own preferences x = 1, planners’ time preferences can be represented by a
function that is a close approximation to a discounted utilitarian IWF, but still assigns
non-zero weight to all future selves.
To calibrate the values of γi, αi in (22), I use the fact that when x = 1 the model
reduces to a set of N independent intertemporal preferences of the form:
V iτ = U i(cτ ) + γi
∞∑s=1
(αi)sV i
τ+s, (A.34)
where αi ∈ (0, 1) and γi ∈ (0, 1−αiαi
). These time preferences have been studied by Saez-
Marti & Weibull (2005), and axiomatized by Galperti & Strulovici (2017). It is straight-
forward to show that they have the following equivalent representation:
V iτ = U i(cτ ) +
∞∑s=1
κsi
(1 + γiγi
)s−1
U i(cτ+s), where κi = αiγi. (A.35)
20
0 1 2 3 4 5 6 7 8
Pure Rate of Social Time Preference %/yr (ρi)
0
1
2
3
4
5
6
Ela
sticity o
f M
arg
ina
l U
tilit
y (η
i)
Figure F.2: Experts’ recommended values for the pure rate of social time preference (ρi),and the elasticity of marginal utility (ηi) for appraisal of long-run public projects, from theDrupp et. al. (2018) survey. 173 responses were recorded. The dashed box depicts datapoints that fall inside the 5− 95% ranges of both parameters. The red cross indicates thelocation of the median values of ρi and ηi.
Writing out the sequence of intertemporal utility weights in this model explicitly,
1, κi,
(1 + γiγi
)κ2i ,
(1 + γiγi
)2
κ3i ,
(1 + γiγi
)3
κ4i , . . . , (A.36)
it is clear that if we take the limit as γi → ∞ of this model holding κi fixed, we recover
discounted utilitarian time preferences with discount factor κi. For any finite γi the pref-
erences in (A.35) are quasi-hyperbolic, with a short run pure time discount factor given by
κi, and a long-run pure time discount factor given by(
1+γiγi
)κi.
Recall that the data in Figure F.2 correspond to the parameters of a discounted util-
itarian IWF, and that our calibration assumption is that these data correspond to the
21
x→ 1 limit of the non-dogmatic model (22). The sequence in (A.36) shows that to ensure
consistency with the calibration assumption we must calibrate κi so that
κi = e−ρi , (A.37)
where ρi is survey respondent i’s recommended value for the pure rate of social time prefer-
ence. In addition, we must choose γi sufficiently large that the model closely approximates
discounted utilitarian time preferences. Notice from (A.36) that the discount factor of
planner i for s > 1 is given by
(1 + γ−1i )κi ≈ e−(γ−1
i +ρi)
when γ−1i is small. Thus γ−1
i = 1%, for example, corresponds to an additional 1%/yr
discount rate on the long-run future, over and above the short run discount rate ρi. Thus
if γ−1i is too large, the model will provide a poor fit to a discounted utilitarian IWF when
x = 1, since non-dogmatic planners will exhibit sharply quasi-hyperbolic time preferences
in this case. To ensure that the model is a close approximation to discounted utilitarianism
when x = 1, but also that all planners place non-zero weight on all future selves’ IWFs
(which requires γi be finite), we must pick γ−1i to be small but non-zero for all i, i.e.,
γ−1i ≈ 0.1%. The numerical results presented in the paper are robust to heterogeneity in
γ−1i , provided that none of these parameters is too large relative to respondents’ pure rates
of social time preference. As stated, γ−1i must be small if the calibrated model is to provide
a good approximation to discounted utilitarian IWFs at x = 1.
In addition, I assume in line with Drupp et. al. (2018) that planners’ utility functions
are iso-elastic, i.e.,
U i(c) =c1−ηi
1− ηi(A.38)
for some ηi > 0. This implies that the elasticity of marginal utility is constant and equal
to ηi, and I simply calibrate ηi to be each respondent’s preferred value of this elasticity.
The requirement that the calibrated model provide a close approximation to discounted
utilitarian IWFs in an appropriate ‘dogmatic’ limit implies that the results depicted in
Figure 1a are robust to alternative specifications of the weights wijs for s > 1. The reason
for this is that, as discussed above (and as is evident from (A.36)), in order for the model
to closely approximate discounted utilitarian IWFs at x = 1, the calibrated values of γi
must be large, which in turn implies that the values of αi must be correspondingly small
22
since κi = e−ρi = γiαi, where ρi is the observed pure time preference rate recommendation
of respondent i. Now notice that the models in (22) can be written as
V iτ = U i(cτ ) + γi
[αi
N∑j=1
wij1 Vjτ+1 + (αi)
2
N∑j=1
wij2 Vjτ+2 +O((αi)
3)
].
Since (αi)s � αi for all s ≥ 2 if αi � 1, it does not much matter how the weights wijs
behave for s ≥ 2. Even if a weight x is given to current preferences at every future maturity,
i.e.,
wijs =
{x i = j
1−x1−N i 6= j
(A.39)
for all s ≥ 1, the results of the simulations hardly change.5
H Changing the model’s time step
This section of the appendix describes how to transform the parameters of the model used
in Figure 1 when the time step is changed.
For the version of the model in question planners’ time preferences took the form
V iτ = U i(cτ ) + γi
[αi
N∑j=1
(P)i,jVjτ+1 + (αi)
2
N∑j=1
(P2)i,jVjτ+2 +O((αi)
3)
]
where P is the annual transition probability matrix defined in (22), which depends on the
parameter x, i.e., the chance of a preference change in a year.
If the model’s time step is changed from 1 year to ∆T > 0 years the values of all its
dynamical parameters must change as well. Consumption growth rates are multiplied by
∆T , and, as in the calibration methodology set out in Section G above, the values of αi
and γi must be recalibrated so that:
κi = αiγi = e−ρi∆T ,
γ−1i ≈ 0.1%×∆T
Transforming the matrix P is more complex. To make the version of the model with time
5Planners with beliefs (A.39) do not obey the consistency condition (14), but this has no relevance forthis discussion.
23
step ∆T comparable to the original annual model, we need to find a stochastic matrix Q
such that
Q = P∆T . (A.40)
When ∆T is not a positive integer (e.g., if ∆T = 1/12 for a monthly time step) such
matrix equations may have no solution, or multiple non-negative solutions. However, in
our case the structure of the model ensures that there is a natural ‘∆T th power’ of P for
any ∆T > 0, and for all interesting values of the parameter x.
Begin by observing that the eigenvalues of P are 1 (with algebraic multiplicity 1)
and Nx−1N−1
(with algebraic multiplicity N − 1), and are thus positive provided that x >
1/N .6 Matrices with positive eigenvalues have a unique ‘principal power’ that satisfies
the equation (A.40) and itself has positive eigenvalues (see e.g., Horn & Johnson, 2013).
It is essential that transforming the time step of the model does not change the signs of
the eigenvalues of the model’s transition probability matrix. If this were not the case the
qualitative dynamics of preference change would not be preserved under a change of time
step. One could, for example, find that planner’s intratemporal weights wijs oscillate with
maturity s, where no such behaviour existed before.
Since P is diagonalizable, it can be written as
P = VDV−1
where
V =
1 −1 −1 . . . −1
1 1 0 . . . 0
1 0 1 . . . 0...
......
. . ....
1 0 0 . . . 1
is a matrix whose jth column corresponds to the jth eigenvector of P, and D is a diagonal
matrix of corresponding eigenvalues, i.e., (D)1,1 = 1, (D)j,j = Nx−1N−1
for j 6= 1. The principal
∆T th power of P is given by
Q = VD∆TV−1.
for any ∆T > 0.
Consider the case ∆T = 1/12, corresponding to a model with a monthly time step. It
6The case x < 1/N is not plausible.
24
0 20 40 60 80 100 120 140 160 180 200
Maturity s (yrs)
0
1
2
3
4
5
6
7
5-9
5%
ra
ng
e o
f r
i (s)
(%/y
r)
x = 80% (annual)
x = 90% (annual)
x = 95% (annual)
x = 97.5% (annual)
x = 1
Figure F.3: Replication of Figure 1a in the paper for a monthly time step. To facili-tate comparison with Figure 1a monthly discount rates have been converted to annualequivalents (vertical axis), and the horizontal axis is scaled to years, rather than months.
is clear from the definition in the previous equation that raising Q to the twelfth power
yields the original matrix P, and that Q has positive eigenvalues. The matrix Q is the
only 12th root of P that has these properties.7
Figure F.3 presents an analogue of Figure 1a in the paper, however this time I have
calibrated the model with a monthly time step using the procedure outlined above. The
figure shows that there is no appreciable difference between versions of the model defined at
different time steps provided that the model parameters are adjusted to reflect the change
in time step.
Finally, I note that any version of the model defined with a discrete time step can be
7Other solutions of (A.40) have the same basic form as Q however we may replace any of the entries onthe diagonal of D1/12 with any of the twelve complex roots of the corresponding eigenvalue of P. As thereis only one way of choosing these roots so that they are all positive (and real), there is a unique ‘principalpower’ of P.
25
thought of as an approximation to an underlying continuous model. Preferences could
change at any instant, and there is some underlying infinitesimal transition probability
matrix that could describe this continuous Markov process. But any discrete approximation
of this process, at any temporal resolution, is legitimate – any behaviour of the continuous
process, when aggregated up to a discrete time step ∆T by exponentiating the infinitesimal
transition matrix, can be replicated by an ‘ab initio’ discrete model with time step ∆T .
We lose nothing (at resolution ∆T ) in this discrete approximation, although the entries of
the discrete transition probability matrix (and hence the weight x) will differ according to
the magnitude of ∆T .
I Decomposing non-dogmatic SDRs
This section studies the resolution of disagreement about the two components of the SDR
– pure time preference and the consumption growth/inequality aversion term – separately.
It shows that much of the rapid convergence of SDRs with maturity shown in Figure 1a is
due to exponential convergence in the consumption growth term.
Section C of the appendix showed that the set of IWFs consistent with (12) can be
represented by
V iτ =
∞∑s=0
N∑j=1
aijs Uj(cτ+s),
where the coefficients aijs are determined by the difference equations in (A.11), and aii0 =
1, aij0 = 0 if i 6= j. Planner i’s SDR at maturity s is
ri(s) = −1
sln
(∑Nj=1 a
ijs (U j)′(cτ+s)
(U i)′(cτ )
)
We decompose this expression into a pure time preference term and a consumption growth
term. Defining
ρi(s) = −1
sln
(N∑j=1
aijs
)(A.41)
Gi(s) = −1
sln
∑Nj=1 a
ijs (U j)′(cτ+s)(∑N
j=1 aijs
)(U i)′(cτ )
. (A.42)
26
we have
ri(s) = ρi(s) +Gi(s). (A.43)
To understand the meaning of ρi(s), notice that∑N
j=1 aijs is the total weight on utilities at
maturity s in IWF i, i.e., it is a pure time discount factor. Hence ρi(s) is IWF i’s pure
rate of social time preference at maturity s. To interpret Gi(s) it is helpful to consider the
case where the utility functions U i(c) are iso-elastic as in (A.38). Denoting the compound
annual consumption growth rate at maturity s by gs, we have8
Gi(s) = −1
sln
(∑Nj=1 a
ijs e−ηjgss∑N
j=1 aijs
). (A.44)
Consider a hypothetical case in which planners have no normative insecurity, i.e., aijs = 0
for all j 6= i; in this case we see that Gi(s) = ηigs, and we recover the familiar consumption
growth term in the Ramsey rule. Gi(s) is the generalization of this term to the non-
dogmatic case, i.e., it is the contribution to the discount rate from consumption growth
and inequality aversion. Figure F.4 plots the range of values for ρ(s) and G(s) as a function
of maturity for the model calibration described in Section G of the appendix. The figure
shows two important things. First, disagreements about the consumption growth term are
significantly larger, and thus quantitatively more important, than disagreements about the
pure rate of social time preference.9 Second, although the range of values for G(0) is larger
than that for ρ(0), disagreements about this term reduce substantially faster as maturity
s increases. The expression for Gi(s) in (A.44) suggests why this occurs. The argument of
the log in this expression is a weighted sum of exponential functions, and thus converges
exponentially fast to e−minj{ηjgs}s as s increases. For example, if consumption growth is a
constant 2%/yr, and we take η = 2 as a modal value of η, and η = 0.05 as the smallest
value of η, at a maturity of 50 years we have e−2×0.02×50 = 0.13, and e−0.05×0.02×50 = 0.95.
Thus values of ηigs that differ substantially from minj{ηjgs} receive little weight at long
maturities, causing the values of G(s) to converge rapidly.
To relate variation in the components ρ(s) and G(s) back to variation in the SDR
8For convenience in this calculation we have chosen units so that current consumption cτ = 1. This iswithout loss of generality.
9The reader may wonder why the ranges for ρ(s) and G(s) depicted in Figure F.4 do not sum to therange for r(s) in Figure 1a. The answer is that the ranges in Figure F.4 are properties of the marginaldistributions of ρ(s) and G(s), while the range of their sum r(s) depends on the joint distribution of thesetwo quantities. Figure F.4 demonstrates how disagreements about these two independently meaningfulquantities reduce as a function of maturity.
27
05
01
00
15
02
00
Ma
turity
s (
yrs
)
0123456 5-95% range of ρi(s) (%/yr)
x =
80
%
x =
90
%
x =
95
%
x =
97
.5%
x =
1
05
01
00
15
02
00
Ma
turity
s (
yrs
)
0123456
5-95% range of Gi(s) (%/yr)
x =
80
%
x =
90
%
x =
95
%
x =
97
.5%
x =
1
Fig
ure
F.4
:R
ange
ofva
lues
for
the
two
com
pon
ents
ofth
eSD
R–
the
pure
rate
ofso
cial
tim
epre
fere
nce
(ρ(s
),le
ft),
and
the
consu
mpti
ongr
owth
term
(G(s
),ri
ght)
–as
afu
nct
ion
ofm
aturi
tys.
The
model
calibra
tion
isth
esa
me
asin
Fig
ure
1a.
28
r(s) = ρ(s) +G(s), we make use of the fact that
Var r(s) = Var ρ(s) + Var G(s) + 2Cov{ρ(s), G(s)}. (A.45)
Figure F.5a breaks the total variance in r(s) into each of these three components at each
maturity, for the illustrative case x = 97.5%. This figure confirms that much of the
variation in r(0) derives from variation in the growth term G(0), but that as maturities
increase disagreements about this term rapidly evaporate. Figure F.5b plots the ratioVarρ(s)
Varr(s) as a function of s for a range of values of x, showing that for all these parameter
values almost all the remaining variation in r(s) for s > 50 is attributable to variation in
ρ(s) – we have almost complete convergence on the dominant G(s) term at these maturities.
29
20 40 60 80 100 120 140 160 180 200
Maturity s (yrs)
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5V
ar
r(s)
×10-4
Var ρ(s)
Var G(s)
2Cov(ρ(s),G(s))
(a) Components of the variance of r(s) (see equation (A.45)). x = 97.5%.
20 40 60 80 100 120 140 160 180 200
Maturity s (yrs)
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
Va
rρ(s
) /
Va
r r(
s)
x = 80%
x = 90%
x = 95%
x = 97.5%
x = 1
(b) Share of the variance of r(s) due to the variance of ρ(s).
Figure F.5: Decomposition of the variance of r(s).
J References
Bergstrom, Theodore C. 1999. “Systems of Benevolent Utility Functions.” Journal
of Public Economic Theory, 1(1): 71–100.
Drupp, Moritz, Mark C. Freeman, Ben Groom, and Frikk Nesje. 2018. “Dis-
counting Disentangled.” American Economic Journal: Economic Policy, 10(4): 109–
134.
Duchin F., Steenge B. 2009. “Mathematical Models in Input-Output Economics”, in
W. Zhang ed., Mathematical Science, UNESCO Encyclopedia of Life Support Sys-
tems.
Friedland, Shmuel. 1981. “Convex spectral functions.” Linear and Multilinear Algebra,
9(4): 299–316.
Galperti, Simone, and Bruno Strulovici. 2017. “A Theory of Intergenerational
Altruism.” Econometrica, 85(4): 1175–1218.
Horn, Roger A., Charles R. Johnson 2013. “Matrix Analysis.” Cambridge University
Press, New York, 2nd edition.
Kingman, J.F.C. 1961. “A convexity property of positive matrices.” Q J Math,
12(1): 283–284.
Marshall, Albert W., Ingram Olkin and Barry C. Arnold 2010. “Inequalities:
Theory of Majorization and its applications.” Springer, New York, 2nd edition.
Nussbaum, Roger D., 1986. “Convexity and log-convexity for the spectral radius.”
Linear algebra and its applications, 73: 59–122.
Prelec, Drazen. 2004. “Decreasing Impatience: A Criterion for Non-stationary Time
Preference and “Hyperbolic” Discounting.” Scandinavian Journal of Economics,
106(3): 511–532.
Saez-Marti, Maria, and Jorgen W. Weibull. 2005. “Discounting and altruism to
future decision-makers.” Journal of Economic Theory, 122(2): 254–266.
Sternberg, Shlomo. 2014. Dynamical Systems. Dover Publications.
31